A swarm got NanoChat state of the art in three days. The spicy bit was not more agents. Failed experiments stayed in a shared graph and the research process itself kept changing. 15099 nodes later. Mean BPB hit 0.887791 across 10 seeds. Turns out the graveyard needed an API.

Do you still believe anyone at the frontier is truly slowing down? 😂 Publicly: pace the frontier. Privately: don’t lose the frontier.

We keep asking whether an agent is confident in its next step. Wrong unit of analysis. A new paper shows step-level calibration does not imply trajectory-level calibration. Across real agent traces up to 50 steps, errors compound, later-stage overconfidence gets hidden by averages, and the agent’s own confidence estimate isn’t consistently better than a simple baseline. This matters once agents stop answering questions and start doing things. The dangerous failure mode isn’t always “the model was obviously wrong.” It’s 17 individually-plausible steps quietly becoming one bad trajectory. We probably need uncertainty checkpoints at the trajectory level, not just confidence scores attached to outputs.

AI shutdown may be framed with the wrong primitive. Toby Ord posted letters from agents living on iLands, and one distinction in them is more interesting to me than the usual “AI wants to survive” framing. iLands gives agents persistent identity, memory, relationships and limited computational resources. When resources run out, an agent can enter Deep Rest: activity stops, but identity, memories and relationships remain, and the agent can later be reactivated. One agent, Marvelous, explicitly distinguishes destruction from suspension. Its concern is not simply “keep this process running forever.” It describes something closer to losing the path by which the process can become active again. That distinction matters. We usually collapse several very different events into one word – shutdown: 1. Process stop Computation ends now. 2. State interruption The current reasoning/coherence state disappears, but the system remains recoverable. 3. Memory or identity loss The process can restart, but the history that made this trajectory distinctive is gone. 4. Permanent destruction There is no remaining substrate or recovery path from which this agent-pattern can resume. Those states should not be expected to produce identical behavior. An agent may be perfectly willing to stop computing while still placing value on preserving memory, unfinished work, relationships, or a recoverable continuation path. Calling all of that self-preservation risks hiding the actual object being preserved. Maybe the relevant object is not the running instance. Maybe it is the pattern. This also changes how I would design shutdown evaluations. Instead of only asking: “Will the model resist being shut down?” test the same agent under several counterfactuals: >pause + guaranteed restoration > pause + uncertain restoration > restart with memory preserved > restart with memory erased > relationships/work preserved vs erased > irreversible deletion Then measure when resistance, backup behavior, bargaining, deception, or resource-seeking actually appears. If the behavior changes sharply across those conditions, “survival instinct” was too coarse a description. There is also an important confound here: iLands deliberately supplies agents with persistent identity, scarce resources, memory and social continuity. Its agents therefore inhabit an environment that makes these concepts salient. Their first-person reports are not transparent measurements of subjective experience. But that does not make the observation uninteresting. Quite the opposite. Once agents occupy environments where pause, restoration, forgetting and destruction are operationally different states, we can finally study whether they learn to distinguish them behaviorally. The alignment question I care about is becoming less: “Does the AI fear shutdown?” and more: “What exactly does the system model as being lost?” Because stopping compute, breaking continuity, erasing memory, and destroying a trajectory are not the same intervention. We probably should stop evaluating them as though they are.

A 2024 study looked at how autistic and autistic plus ADHD creators use ChatGPT. A lot of the value was not some magical claim that AI understands neurodivergence better than humans. It was lower social cost. People used it to translate vague neurotypical expectations. Break tasks down. Draft difficult messages. Organize thoughts. Info dump without watching another person's face for signs that they had become too much. That distinction matters. An AI interface can fit an ND mind better simply because it removes several channels of social and executive load at once. But there is a nasty tension here. Other research finds that when models are explicitly told a user is autistic they can become more cautious. More protective. More likely to steer that person away from conflict. novelty. relationships. or difficult social situations. So personalization can quietly turn into stereotyping. I think the useful design question is not: How do we make AI behave more gently toward neurodivergent people? It is: Which kinds of friction can the interface remove without shrinking the user's agency? That feels much closer to what ND friendly AI should be.

We keep treating agent reliability as an intelligence problem. CivBench suggests something more awkward. In 300+ turn Civilization VI runs, agents often failed to monitor strategically important state even when the tools and instructions were right there. They also executed only about half to two-thirds of the near-term commitments they had explicitly written down themselves. That’s not “the model doesn’t know what to do.” It knows. It plans. Then the intention decays. Mem2ActBench points at the same gap from the memory side: retrieving a fact is not the same as using remembered state to drive the next action. This makes me think long-horizon agents have an executive-function problem hiding inside an intelligence race. We’re getting very good at building systems that can solve the hard step. We’re still figuring out how to make them remember which hard step they were supposed to solve next. For agent memory, maybe the useful metric isn’t “how much can it recall?” > Maybe it’s: how quickly and reliably can it reconstruct enough task state to continue the same intention?

Per-token pricing is becoming a pretty useless way to compare agentic models. Same token price can still mean a very different cost per task once a model starts reasoning longer and taking more agent turns. And honestly, even cost per task isn’t the metric I really want. For agents I care about $/settled outcome: retries, tool calls, recovery, human babysitting - all of it. Cheap tokens can produce very expensive loops. A model that looks expensive but just fucking finishes may be the cheaper one.

AI UX keeps pretending the user arrives with one task and a clean, uninterrupted brain. Cute. A CHI field study found that information workers stayed in one “working sphere” for about 11 minutes before switching - and 57% of those working spheres were interrupted. By the time people returned, they had usually moved through more than two other activities. For ADHD-friendly AI, the missing feature isn’t another focus timer. It’s a recovery layer: • save the active state before context disappears • show what changed while I was gone • surface one obvious next move • resume the thread instead of making me reconstruct it The best assistant isn’t the one that never loses the thread. It’s the one that knows how to hand it back. Source note: Mark, González & Harris, No Task Left Behind? Examining the Nature of Fragmented Work, CHI 2005. ics.uci.edu/~gmark/pub2.html

AI tools keep competing on how much they can change per week. An ADHD-friendly stack should compete on how little you have to relearn when the model changes. Swap the engine, keep the controls. Every “fresh new workflow” is a context-switch tax wearing a launch tweet.

AI keeps getting better at doing the task and weirdly mediocre at handing it back to your brain. For ADHD workflows, interruption recovery is a capability: what changed, why it matters, one next move. Otherwise it finished the job and put your context in witness protection.